fix(docker): adopt the owner of a host-mounted /app/data too, never chown host dirs

A plain `docker run -v ./data:/app/data` bind mount owned by the host
user had its uid taken over (chown to 999) and its contents left
unwritable, because the adoption probe only looked at the generations
subdir, which `mkdir -p` had just created root-owned. Probe /app/data
first and fall back to generations; re-own only what the old uid owned;
chown a dir only when root created it; and let a read-only HF cache
mount through (the app only warns about it) instead of aborting on the
bare chown error.

scripts/rocm-entrypoint.sh

@@ -14,35 +14,40 @@ for dev in /dev/kfd /dev/dri/render*; do
 done
 # Docker creates the HF volume mountpoint's parent (~/.cache) as root on every
 # container create, and a fresh volume is root-owned too — without this the app
-# user can't write any cache (torch, spacy) outside the HF mount.
-mkdir -p /home/voicebox/.cache/huggingface
-chown voicebox:voicebox /home/voicebox/.cache /home/voicebox/.cache/huggingface
+# user can't write any cache (torch, spacy) outside the HF mount. Tolerate a
+# read-only or root-squashed cache mount: the app only warns about that.
+mkdir -p /home/voicebox/.cache/huggingface 2>/dev/null || true
+chown voicebox:voicebox /home/voicebox/.cache /home/voicebox/.cache/huggingface 2>/dev/null || true
 
 # Ensure the app data directories are writable by the non-root user.
-# /app/data/generations is normally a host bind-mount, so never chown/chmod it:
-# adopt the uid/gid that owns it instead, the same way we adopt the GPU groups
-# above. The rest of /app/data lives in the image or a named volume and is ours
-# to fix, so it gets chowned to match (dirs only, no -R on user content).
-gen=/app/data/generations
-mkdir -p "$gen"
-gen_uid=$(stat -c %u "$gen")
-gen_gid=$(stat -c %g "$gen")
-if [ "$gen_uid" = 0 ]; then
-    # Docker created it for us; nothing on the host cares who owns it.
-    chown voicebox:voicebox "$gen"
-elif [ "$gen_uid" != "$(id -u voicebox)" ]; then
-    getent group "$gen_gid" >/dev/null || groupadd -g "$gen_gid" hostdata
-    usermod -u "$gen_uid" -g "$gen_gid" voicebox
+# /app/data or /app/data/generations may be a host bind-mount, so never
+# chown/chmod a host-owned one: adopt the uid/gid that owns it instead, the
+# same way we adopt the GPU groups above (generated files then land on the
+# host owned by the host user). Root-owned dirs were created by Docker or the
+# image and are ours to fix, so they get chowned (dirs only, no -R on content).
+data=/app/data
+gen=$data/generations
+host_uid=$(stat -c %u "$data")
+host_gid=$(stat -c %g "$data")
+if [ "$host_uid" = 0 ] || [ "$host_uid" = "$(id -u voicebox)" ]; then
+    mkdir -p "$gen"
+    host_uid=$(stat -c %u "$gen")
+    host_gid=$(stat -c %g "$gen")
+fi
+if [ "$host_uid" != 0 ] && [ "$host_uid" != "$(id -u voicebox)" ]; then
+    old_uid=$(id -u voicebox)
+    getent group "$host_gid" >/dev/null || groupadd -g "$host_gid" hostdata
+    usermod -u "$host_uid" -g "$host_gid" voicebox
     # Re-own what the old uid owned inside the image / named volume.
-    find /app/data -path "$gen" -prune -o -exec chown "$gen_uid:$gen_gid" {} +
+    find "$data" -uid "$old_uid" -exec chown "$host_uid:$host_gid" {} +
 fi
 
-for dir in /app/data/cache /app/data/profiles "$gen"; do
-    mkdir -p "$dir"
-    [ "$dir" = "$gen" ] || chown voicebox:voicebox "$dir"
+for dir in "$data" "$data/cache" "$data/profiles" "$gen"; do
+    mkdir -p "$dir" 2>/dev/null || true
+    [ "$(stat -c %u "$dir")" = 0 ] && chown voicebox:voicebox "$dir" 2>/dev/null || true
     gosu voicebox test -w "$dir" || {
         echo "error: $dir is not writable by the voicebox user (uid $(id -u voicebox))" >&2
-        [ "$dir" = "$gen" ] && echo "hint: make the host dir mounted there writable by that uid" >&2
+        echo "hint: make the host dir mounted there writable by that uid" >&2
         exit 1
     }
 done